Research map / Adversarial robustness and security / Attacks on text and graphs
On the Robustness of Semantic Segmentation Models to Adversarial Attacks
Anurag Arnab, Ondrej Miksik, Philip H.S. Torr
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a ResNet backbone instead of VGG | FCN8s and Deeplab v2 ASPP · Pascal VOC | adversarial robustness | Improved |
| Use Deeplab v2 with Multiscale ASPP instead of its single-scale variant (ablation) | Deeplab v2 · Pascal VOC and Cityscapes | adversarial robustness | Improved |
| Add the Dilated-Net Context module (ablation) | Dilated-Net Front-end · Cityscapes and Pascal VOC | adversarial robustness | Hurt |
| Use a lightweight residual architecture rather than a chain-like VGG-based network | Semantic segmentation models · Pascal VOC and Cityscapes | adversarial robustness | Improved |
| Process inputs at multiple scales (ablation) | Deeplab v2 · Pascal VOC | IoU ratio under adversarial attacks | Improved |
| Generate perturbations at one scale and evaluate them at another | Deeplab v2 and other segmentation networks · Pascal VOC | attack effectiveness | Hurt |
| Add mean-field CRF inference (ablation) | FCN8s · Pascal VOC, untargeted attacks | adversarial robustness | Improved |
| Add mean-field CRF inference (ablation) | FCN8s · Pascal VOC, targeted attacks | adversarial robustness | No clear effect |
| Generate black-box perturbations from the FCN8s unary output rather than CRF-RNN output | CRF-RNN · Pascal VOC | attack effectiveness | Improved |
| Increase CRF pairwise-potential bandwidth to produce more smoothing (ablation) | CRF-RNN · Pascal VOC | adversarial robustness | No clear effect |
| Use single-step FGSM on Cityscapes rather than Pascal VOC | Segmentation models · Cityscapes versus Pascal VOC | IoU ratio under FGSM: 2.5% versus 27.9% for PSPNet; 2.8% versus 12.2% for Dilated Context | Hurt |
| Use single-step FGSM rather than targeted Iterative FGSM ll at high perturbation levels | Segmentation models · Cityscapes | attack effectiveness | Improved |
Closest papers by meaning
- Towards Reliable Evaluation and Fast Training of Robust Semantic Segmentation Models (ECCV 2024 · similarity 0.94)
- Evaluating the Adversarial Robustness of Semantic Segmentation: Trying Harder Pays Off (ECCV 2024 · similarity 0.94)
- Characterizing Adversarial Examples Based on Spatial Consistency Information for Semantic Segmentation (ECCV 2018 · similarity 0.92)
- Dynamic Divide-and-Conquer Adversarial Training for Robust Semantic Segmentation (ICCV 2021 · similarity 0.92)
- RP-PGD: Boosting Segmentation Robustness with a Region-and-Prototype Based Adversarial Attack (AAAI 2025 · similarity 0.89)
- AAKR: Adversarial Attack-based Knowledge Retention for Continual Semantic Segmentation (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
